(Deno) MiniMax H3 Reference to Video

If you've ever stared at the stock MiniMax H3 "Reference to Video" node and wondered which of the half-dozen image slots to wire into where - this is the version that answers that. (Deno) MiniMax H3 Reference to Video is a thin wrapper around ComfyUI's native H3 node that swaps the Autogrow image-input mess for a single ordered reference bundle. Everything else stays stock, including the prompt tags.

Context first, because it's why this node exists. MiniMax H3 (the open-weight model MiniMax/Hailuo announced in July 2026) is multimodal: it takes text, reference images, reference videos, and reference audio, and generates video with native sound - up to 15 seconds at 2K in the hosted tier, with reference inputs "baked in from the start," as the community put it. ComfyUI shipped native H3 nodes in 0.30.0, and the stock reference-image handling is a set of Autogrow slots that break on mixed-size references. Deno's loader + this node replace that with one cable.

How it works

The node subclasses ComfyUI's own MiniMaxH3ReferenceToVideo, so the mechanism is stock H3 - it's not a reimplementation. It re-exposes the ref_images input as the custom DENO_MINIMAX_H3_REFERENCE_IMAGES type from (Deno) MiniMax H3 Multi Reference Image Loader, and at execution it unpacks that ordered bundle back into the per-slot images H3 actually wants. The ref_videos, ref_video_audios, and ref_audios inputs stay as stock Autogrow sockets, in prompt-tag order. Because the image bundle is order-preserving, <Picture 1>, <Picture 2>, and so on in your prompt map straight to your loader's card order - same for <Video k> and <Audio j>.

Inputs that matter

• clip, vae, audio_vae - H3's text/image encoder and its video and audio VAEs, straight from the stock MiniMax H3 loaders.
• prompt - your generation prompt; reference images are <Picture 1>, <Picture 2>..., videos <Video 1>..., audio <Audio 1>...
• width / height - 1344×768 by default.
• length - frame count at 24 fps; default 124 (about 5 seconds), steps of 17.
• ref_image_size - match (default) sizes references to the generation area; the other option runs the larger max-reference pipeline.
• Optional: ref_images (the DENO bundle), plus stock ref_videos, ref_video_audios, ref_audios.

Outputs are the standard H3 pair: positive CONDITIONING with all the image/video/audio references baked in, and an empty LATENT for the sampler. Wire those into a MiniMax H3 sampler and decode.

Install

Part of the Deno Custom Nodes pack - ComfyUI Manager, search "deno custom nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/Deno2026/comfyui-deno-custom-nodes.git
cd comfyui-deno-custom-nodes
python -m pip install -r requirements.txt

Restart, then make sure you're on ComfyUI 0.30.0 or newer - the node literally cannot exist without the native H3 node underneath. The H3 model files (checkpoint, CLIP, video VAE, audio VAE) load through the stock MiniMax H3 loaders in the same workflow; the Deno pack doesn't bundle or download them.

Where people get burned

The bundle is enforced, not optional: connect anything other than the DENO loader's ref_images and you'll get a clear "ref_images must come from the DENO MiniMax H3 Reference Image Loader" error. Empty bundle, same story. If the node itself is missing from the menu, your ComfyUI predates 0.30.0. And a reality check on expectations: H3 is a heavy multimodal model - images, video, and audio references plus generation on one card adds up, and reference images can be downscaled internally by ref_image_size regardless of what you feed in. Test one reference first, then scale up. The loader keeps your source pristine; the model's own reference-sizing pipeline is the part you tune.